Unfortunately there seem to have been a large number of informal experiments, with various shortcomings such as using a limited range of source material, focusing on preference without also looking at ability to tell a difference, or making the files listened to available to test subjects (removing some or all of the 'blind' quality from the test) - but relatively few published 'scientific' studies.
Most lossy compression algorithms I know of are based on perceptual coding - a nicely detailed paper describing some of the principles (in the context of the development of PASC, an early technique) is at http://www.minidisc.org/MaskingPaper.html.
A few points from that paper :
the ear divides the sounds it receives into frequency bands, called 'critical bands'. Quiet sounds in the critical bands can be 'masked' by other, louder sounds in the same critical band. (this gives the opportunity for data reduction : if we can't hear a quieter sound, we don't need to encode it).
"There is no agreement as to the specific number of critical bands active simultaneously. Critical bands and their center frequencies are continuous, as opposed to having strict boundaries at specific frequency locations".
we have many more bands in the lower frequency range
In general, low sounds mask higher sounds
Sometimes a signal can be masked by a sound preceding it, called forward masking, or even by a sound following it, called backward masking
The masking threshold of a signal entering in one ear can be raised by a masker entering in the other ear.
Spatial location can have a negative effect on masking
Thinking about all those in combination, it's clear that the characteristics of the audio we're encoding are going to have a huge influence on the amount of data we can throw away and still keep the result subjectively indistinguishable from the original (or at least, not subjectively worse). So, even taking account file format and compression algorithm, another huge variable is the song we're encoding.
This means that as a general rule, taking a rough average of the results from the 'unscientific' tests we've read about may give as good an answer to your question as we can hope for without considering a particular song.
For what it's worth, my personal impressions, for music, on average:
- 128Kbps is usually just about good enough to enjoy the song, but is almost always noticeably worse than the original
- 192Kbps is always good enough to enjoy the song, and sometimes be indistinguishable from the original (or at least, not subjectively worse).
- 320 Kbps is close enough that I'm not putting any money on that I can distinguish it from the original in most cases; (but I would bet I can find some particular tracks where I can).
On yet another personal note : I listen to music through playlists, so I'm not usually aware what file format or bitrate I'm listening to. Sometimes I catch a bit of a song and I think "wow, that sounds good!" and I check the file - very often (85% of the time?) it is lossless or 320kbps; just occasionally it's 192. It's never less.
